Words? |? Yu yue
Last weekend, China 2020, which was delayed for more than half a year, officially opened. Due to the need of epidemic prevention and control, Beijing Auto Show may also be the largest auto show in the world this year.
At this year's China, one of the biggest attractions is that L2 autopilot has gradually become the standard for new cars, and its own brand L3 or even L4 autopilot is gradually landing; Huawei and other technology companies have come up with an autonomous driving solution to "show muscles"; Many self-driving chip companies have proposed explosive solutions ... to make this year's China Auto Show look like an auto technology show.
Black Sesame, a domestic autopilot chip startup, exhibited the FAD autopilot computing platform for the first time, with a computing power of 280TOPS, even surpassing Tesla's self-developed FSD autopilot computer in terms of data.
So, is the black sesame FAD autopilot platform really that strong? How does FAD crush Tesla FSD?
1. The strongest autopilot computing platform at Beijing Auto Show? The computing power can reach up to 280TOPS.
On the first day of the auto show, the domestic self-driving chip company Black Sesame officially released and exhibited FAD (all? Unconscious? Driving) fully automatic driving computing platform, and also exhibited the second automatic driving chip Huashan II A 1000 series for the first time.
The biggest advantage of FAD is its powerful computing power. From the data point of view, it has even surpassed Tesla's strongest autopilot computing computer hardware? 3 (the actual computing power is about 72TOPS), the maximum computing power of a single controller can reach 280TOPS, and the high computing power also enables FAD to support L2+~ L4 autopilot.
▲ Black Sesame FAD autopilot platform exhibited at Beijing Auto Show.
In June this year, Black Sesame officially released Huashan II A 1000 series chips. The maximum computing power of a single chip can reach 70TOPS, and the power consumption is less than 8W. A single chip can support L3 autopilot system. The FAD autopilot platform is based on Huashan II A 1000 chip. A single controller can integrate up to four Huashan II A 1000 chips, and the maximum computing power can reach 280TOPS.
Even with such a powerful computing power, the maximum power consumption of the FAD autopilot platform is tens of watts, and the energy efficiency ratio reaches 6 tops/w.
Such performance data and energy efficiency ratio data are in a leading position even in the field of global autonomous driving chips.
In addition, in order to adapt to different automobile manufacturers and models, automobile enterprises can also choose 1, 2 or 4 domain controllers integrated with Huashan No.2 A 1000 chip, which can realize the maximum computing power of 70TOPS, 140TOPS and 280TOPS respectively, corresponding to L2+, L3, L3/L4 autopilot systems.
In fact, this is to meet the cost needs of car companies.
It is understood that the FAD autopilot platform supports a variety of sensor access, including lidar and millimeter-wave radar, and simultaneously accesses multiple cameras. It also has built-in GNSS+IMU+RTK combined high-precision positioning module, which can realize lane-level positioning and realize more accurate automatic driving.
It can be said that the black sesame FAD autopilot platform has become one of the most powerful autopilot controllers in China with its high computing power and low power consumption.
Second, support L2+~ L4 autopilot? Can black sesame FAD really beat Tesla FSD?
FAD not only uses super chips, but also has super computing power. What other powerful skills does FAD have?
Yesterday morning, Che Dongxi and Black Sesame founder and CEO Shan conducted an in-depth interview with the founder and found the answer to the question.
1. Benchmark Tesla and NVIDIA? Overall leading performance
Among the global manufacturers of self-driving chips, only NVIDIA and Tesla can achieve such a high level of computing power as Black Sesame FAD.
In terms of computing power, the computing power of NVIDIA's Xavier computing platform is 30TOPS, and Tesla's FSD computing platform (hardware? 3) Computing power 144TOPS, a black sesame FAD computing platform integrating two Huashan No.2 A 1000 chips, with the highest computing power 140TOPS.
▲ Black Sesame FAD autopilot platform exhibited at Beijing Auto Show.
On the utilization rate of computing power, the gap between them is quite large. The FAD utilization rate of black sesame can reach over 80%, that of Tesla is 55%, and that of NVIDIA's Xavier computing platform is about 20%.
Shan Zhangzhang said that the utilization rate of computing power is too low when using CPU and GPU for AI operation, and special neural network processors must be used for AI operation to make the utilization rate of computing power reach a relatively high level.
In fact, this is one of the reasons why Tesla gave up NVIDIA autopilot chip and switched to self-developed chip.
In terms of power consumption, NVIDIA Xavier computing platform and black sesame FAD are both around 30W, while Tesla FSD power consumption has more than doubled to 72W. In addition, the energy efficiency ratio of NVIDIA's Xavier computing platform is only 1 tops/W, Tesla is slightly better, with 2TOPS/W, while the black sesame FAD can achieve 6 tops/w. ..
According to reports, although the overall power consumption of the three is relatively low, strict thermal management optimization is also needed. If the energy efficiency ratio is too low, it is difficult to control the chip heating, which may lead to the decline of overall safety and stability.
In May of this year, NVIDIA released an iterative product of autopilot chip based on Ampere architecture. Compared with the previous generation Xavier computing platform, the computing power has been greatly improved, and there are three products: 10TOPS, 200TOPS and 200tops. However, at present, NVIDIA has not released much news about the Ampere architecture autopilot platform, and the mass production time is also scheduled for the second half of 2022, which is still a certain distance from the actual loading. However, Black Sesame predicts that FAD can be mass-produced and loaded in the second half of 202 1.
2. The neural network engine has super performance? Support a variety of artificial intelligence operations
Looking at Huashan II A 1000 chip, we can find that it is mainly composed of an 8-core CPU, GPU, memory controller, neural network processor, digital signal processor (DSP), image signal processor (ISP), security island and I/O interface controller. It is understood that the SoC has 22 processors.
▲ Inside Huashan II A 1000 chip
And so many cores and computing units are integrated in less than 90mm? Area, equivalent to the size of a fingernail. The size of the packaged chip is only about 250mm? .
Among them, DynamAI occupies the largest area? NN Neural Network Processor It is understood that there are four 3D convolution arrays in this area, the main frequency is 1.2GHz, and the computing power can reach 39TOPS；; And 1 2D? GEMM matrix multiplication array, clock frequency is 800MHz. In addition, there are five independent DSPs in the chip, so the computing power of the whole chip can reach 43TOPS. This does not include the computing power of CPU, GPU and CV accelerator.
In fact, this is the most basic computing power of the A 1000 chip of the second generation of Black Sesame Huashan, and the strongest computing power can reach 70TOPS.
At the same time, because there are many different neural network processors, the whole SoC can maintain strong flexibility and balance.
3. lower the threshold for development? Meet the standards of vehicle regulations
At present, most vehicle companies and autonomous driving companies adopt self-developed autonomous driving software and purchase autonomous driving hardware to develop or mass-produce autonomous driving. However, the progress of perception, recognition and decision-making algorithms of each manufacturer is different, and the neural networks used are also different, which requires hardware manufacturers to closely cooperate with the software development process of vehicle enterprises.
Black sesame provides a perfect development solution for different automobile manufacturers and autonomous driving companies.
First of all, Black Sesame provides a self-developed algorithm neural network and a general neural network model base. For example, in the ADAS algorithm neural network, there are networks such as road signs, road signs and lane lines to predict the movement trend of the target. Black Sesame provides these algorithmic neural networks to customers as SDK. You can also make common and open neural networks into model libraries and provide them to customers as open source software.
Second, Black Sesame provides customers with neural network development tools, which can continuously optimize the algorithm. Even before delivery, you can run the developer's network in a virtual environment. Developers can pass the cloud platform test of Black Sesame and don't even need to install it locally.
Third, if customers don't develop autopilot software at all, Black Sesame can also provide hardware, tool chain, software library and operating system for packaging and delivery.
In terms of safety process standards and vehicle code certification, FAD platform conforms to ASIL? ASIL？ Automobile functional safety requirements and CC? Safety certification requirements for EAL5+ vehicles. After upgrading to board-level products, FAD? The dual-chip system of the platform can work completely independently, support the redundant design of power supply and video acquisition system, and meet the design requirements of pre-installed products. Support ASIL through dual chip redundancy? Security decomposition, using ASIL-D level to control MCU, realizing system-level ISO26262? ASIL-D safety process standard. Further meet the safety design requirements of system-level autopilot platform SOTIF and RSS.
It can be said that the black sesame FAD autopilot platform has been ahead of most autopilot chips on the market in terms of hardware computing power, overall efficiency, software development and safety certification, and some parameters have surpassed Tesla FSD at the same level.
Third, industry veterans join hands to build cores? FAD is about to enter the mass production stage
According to the introduction of the Black Sesame Entrepreneurship Team, the company was established on 20 16, and invested in the research and development of autonomous driving chips, which is roughly the same as the start-up time of Tesla's self-developed FSD chip, which makes the two companies stand on the same starting line.
In 20 17 years, weilai hexin kinetic energy invested nearly 100 million yuan in it. From 2065438 to April 2009, Black Sesame received the second round of investment from SAIC, SK China and China Merchants.
These investments have enabled Black Sesame to invest heavily in the research and development of autonomous driving chips. At present, the R&D investment has exceeded 654.38 billion US dollars (more than 680 million yuan). You know, the R&D team of Black Sesame has only 300 people, and they are all top talents from various technical highlands such as chips, automobiles and algorithms.
▲ Left and right are Black Sesame CEO Zhang Zhangzhang and Chief Operating Officer Liu Weihong.
Shan Zhangji, CEO of Black Sesame, was a member of the founding team of OmniVision, the global leader in visual chips. He has worked hard in the silicon valley chip industry for more than 20 years and has rich experience and technical accumulation in image processing chips and software algorithms.
CTO Zheng Qi is the main design member of Intel Pentium II chip, CSO Zeng Daibing is the chief engineer of ZTE Microelectronics, and Chief Operating Officer Liu Weihong is the president of Chassis and Control System Division, the main department of Bosch China ADAS.
It is precisely because of the super R&D team that the startup company Black Sesame can make the ADAS chip Huashan No.1 A500 and mass-produce it within three years. This year, Huashan II A 1000 chip will be launched, and FAD autopilot platform will be released.
Since the beginning of this year, if the new car is not equipped with L 1/L2 autopilot, it will be "embarrassed to sell", and the popularity of autopilot is increasing rapidly. The higher L3 autopilot and even L4 autopilot have reached the eve of mass production, and the industry's demand for autopilot chips and computing platform solutions is exploding. The market scale of autonomous driving chips alone is expected to reach the trillion-dollar level and become the largest single market in the semiconductor industry.
Therefore, it is just the right time for FAD to enter the autonomous driving market.
In August this year, FAW Intelligent Network Development Research Institute reached a technical cooperation agreement with Black Sesame. FAW Intelligent Network Development Research Institute will start the development of intelligent driving platform based on Huashan No.2 A 1000 to meet the demand of subsequent production vehicles. The two sides will jointly promote the application of artificial intelligence technology in the automobile industry and accelerate the industrialization of domestic intelligent driving chips.
In addition, Black Sesame has also signed a number of FAD designated models, and it is expected that there will be models equipped with FAD autopilot platform on the market next year. In addition, many companies at home and abroad have begun to test the FAD autopilot platform, and the test vehicles have been on the road.
The great success of black sesame in autonomous driving chips and domain controllers has made industry research institutions pay attention to this startup that has just been established for more than 4 years. In April this year, CB, one of the strongest think tanks in Silicon Valley? Insights released the list of chip design enterprises in China, and Black Sesame was listed in the field of automobile chips, becoming one of the top 65 chip design enterprises in China.
In July this year, the A 1000 chip of the second generation of Black Sesame Huashan also appeared at the World Artificial Intelligence Conference, together with high-end artificial intelligence chips such as Pingtou Ge, Yi Tu and Cambrian.
It can be said that after more than four years of development, Black Sesame has become the world's leading self-driving chip design company, and even has the ability to compete with the chip industry. At the same time, the rapid progress of black sesame has also promoted the design of domestic autonomous driving chips to a new level.
In the conversation with the two founders, they also revealed an egg. Next year, Black Sesame will release a chip with stronger performance. By then, the maximum computing power of the FAD autopilot platform equipped with this chip is expected to exceed 1000TOPS, and its computing power can be fully automated.
This article comes from car home, the author of the car manufacturer, and does not represent car home's position.